Skip to content

Model Training⚓︎

Scenario: Developing an AI Model

Context⚓︎

We want to be able to train an AI model using the MLOps platform. The project team consists of ML Developers, who will develop the model and train it.

Scenario⚓︎

Step 1: Model Creation⚓︎

  • Actors: ML Developer, GitLab
  • Action: The ML Developer starts a new project on the GitLab platform after login.

Step 2: Editing Project to create metadata and reference in SharingHub⚓︎

  • Actors: ML Developer, GitLab
  • Action: The ML Developer edit the GitLab project to add necessary metadata to the README.md, and add the AI model topic in the project settings, to be referenced in Sharinghub.

Step 3: Training environment⚓︎

  • Actors: ML Developer, MLflow client, SharingHub
  • Action: The ML Developer initialize his training environment with the MLflow client installed, as well as the model dependencies. It can be locally or remotely on a computing platform.

Step 4: Prepare the dataset⚓︎

  • Actors: ML Developer, SharingHub
  • Action: The ML Developer find the dataset on SharingHub, or another source, retrieve it, which may be with Git LFS or DVC, and prepare the dataset for the training.

Step 5: Model Training⚓︎

  • Actors: ML Developer, MLflow client, SharingHub
  • Action: The ML Developer configures with the parameters the model and begins training using the dataset. Metrics are logged with MLflow client, using the tracking URI given by SharingHub.

Step 6: Evaluating Model Performance⚓︎

  • Actors: ML Developer, MLflow UI
  • Action: After training, the ML Developer evaluates the model’s performance using the metrics on the MLflow UI Experiments->runs.

Step 7: Versioning the Model⚓︎

  • Actors: ML Developer, MLflow UI, SharingHub
  • Action: After achieving satisfactory results, the ML Developer save the final model via the MLflow UI in the registry. After that, the model version is automatically linked to the project STAC metadata in SharingHub. This ensures that other team members can test and use the model release.